Personalization for the Grammarly Keyboard

Hasn’t everybody wished, in some unspecified time in the future or one other, that their telephone keyboard understood them just a little higher? We’ve all skilled typing a phrase that we use ceaselessly—our pet’s nickname or the title of a undertaking at work—solely to have our keyboard not acknowledge this time period and provides us an unhelpful correction as a substitute.

In the present day, the Grammarly Keyboard on iOS can study your private lexicon, tailoring strategies to the phrases you employ, even when they’re not in the usual dictionary. On this publish, we’ll clarify how we constructed the mannequin that powers this performance, which runs fully on the system, in a performant, correct method.

Earlier than – Consumer typing their customized vocabulary for the primary time.

After – PLM discovered the consumer vocabulary and auto-corrected phrases.

Why on system?

Whereas personalization can enhance communication, it ought to by no means come on the expense of privateness. At Grammarly, we’re dedicated to making sure that customers at all times management their knowledge. On condition that we’re modeling private vocabulary, we constructed the mannequin fully on the system in order that delicate knowledge would by no means depart the system or be uncovered to 3rd events.

Constructing the mannequin on the system has further advantages when in comparison with a conventional cloud-based mannequin. Because it doesn’t depend on connectivity to operate, customers will at all times have entry to personalised strategies, whatever the stability of their connection. Moreover, we received’t should do any advanced syncing with the server (as within the case of a hybrid mannequin that splits processing between the cloud and the system).

That mentioned, there are a number of challenges with constructing an on-device mannequin, which we’ll focus on subsequent.

Maximizing efficiency

The typical cell system has 4 GB of RAM, of which solely ~70 MB can be utilized by the keyboard at any given time. The Grammarly Keyboard already makes use of 60 MB for core performance, leaving lower than 5 MB for brand new options. What’s extra, keyboard efficiency actually issues—the consumer will acutely really feel any lags when typing.

We do a number of issues to make sure that our personalised mannequin doesn’t decelerate (or crash) the keyboard. First, we retailer the mannequin in persistent reminiscence and use a memory-mapped key-value retailer to retrieve related n-grams into RAM on an on-demand foundation. We additionally cache recurring computations, enabling environment friendly chilly and heat begin occasions. Lastly, we restrict the scale of the variety of unigrams and n-grams within the customized vocabulary dictionary to keep away from bloating the system’s persistent reminiscence storage.

The restricted customized vocabulary dictionary required us to thoughtfully handle the method of including new phrases. Particularly, we would have liked to differentiate between out of date phrases (that we must always neglect) and related phrases (that we must always preserve). We did this by making use of a time-based decay operate that dynamically adjusted phrase chances primarily based on how just lately the phrase was used. When the dictionary will get full, we delete the least-used phrases (as calculated by the operate) to create house for brand new phrases.

Bettering accuracy

Along with efficiency, we targeted on delivering correct strategies. This proved tough, as no reference dictionary exists for a consumer’s private lexicon. Due to this fact, after we encountered a brand new phrase, we would have liked to differentiate if it was legitimate vocabulary or a typo. For instance, let’s say you’re texting hey to your dad, whom you generally seek advice from as “pops.” You’ve typed “heeyyy pops.” Ought to we study heeyyy and pops as new phrases?

We tackled this drawback by first addressing noisy inputs. Noisy inputs are informal variations of precise phrases—they may embody further vowels or consonants to convey tone (“awwwww”), have lacking apostrophes (“cant”), or use incorrect capitalization (“i agree”). We excluded these inputs from our studying course of to satisfy our customers’ expectations for high-quality, skilled strategies. We use a mixture of regex filters and particular guidelines to establish noisy inputs. Solely inputs that aren’t flagged as noisy are discovered by the mannequin. (Within the instance above, we’d categorize heeyyy as informal as a result of further letters, and our mannequin wouldn’t study this enter.)

The earlier strategy doesn’t absolutely handle the query of whether or not to study and recommend phrases like pops. Whereas varied methods exist, we opted for a easy trust-but-verify technique. This includes studying each new phrase however deferring strategies till the phrase seems sufficient occasions. Particularly, we use edit-distance-based frequency thresholding to find out when the candidate has met the required standards to go from studying to suggesting. This technique lets us distinguish high quality new phrases from noise with out requiring costly operations.

To judge the efficacy of our strategy, we constructed an offline analysis framework to simulate manufacturing conduct. This allowed us to validate that the mannequin dealt with potential edge instances correctly and establish potential errors to repair earlier than they affected prospects. In reality, that’s how we found that we weren’t dealing with inputs like “dont” or “cant” correctly, which led to creating new regex filters. Surprisingly, the framework additionally validated that the mannequin did an amazing job studying frequent correct nouns (like iTunes) that weren’t a part of the default dictionary.

Impression

We’ve shipped the personalised mannequin to over 5 million cell units by way of the Grammarly Keyboard. Notably, it’s already having a large constructive influence on our ecosystem.

By our aggregated logging metrics, we’ve noticed a big lower within the fee of reverted strategies and a slight enhance within the fee of accepted strategies. This means that we’re fixing the issue we got down to repair—customers are getting fewer irrelevant strategies that they should revert, which signifies that we’re doing a greater job modeling how they convey. Our inside efficiency metrics additionally present that the mannequin operates with minimal RAM utilization and environment friendly chilly and heat begin occasions, signaling that the keyboard app is responsive.

Takeaways

Creating personalised lexicons on units marks a big milestone in Grammarly’s mission to empower customers to speak extra successfully and prioritize their privateness. By superior methods like adaptive algorithms and good outdated trial and error, we’ve uncovered easy methods to study a consumer’s private lexicon with out requiring cloud computing energy. In case you are enthusiastic about constructing fashions that energy higher digital communication, we’d love to listen to from you. Take a look at our job openings right here.

Particular due to the whole workforce that labored on this undertaking: Sri Malireddi, Suwen Zhu, Kosta Eleftheriou, Dhruv Matani, Roman Tysiachnik, Oleksandr Ivashchenko, Illia Dzivinskyi, Ignat Blazhko, Ankit Garg, John Blatz, and Max Gubin.